Many biases. One transparent score.
TARA listens to performance conversations and surfaces where bias may be creeping in — before a rating locks. Here's what it listens for, and how a single signal becomes a score you can audit.
Why this exists
Bias does its quietest damage in the review
Performance reviews are the moments where months of work collapse into a number. That number shapes pay, promotion, and — over time — whether an employee stays or leaves. It is also, research consistently shows, one of the most bias-saturated moments in a person's working life — not because managers are malicious, but because the whole process depends on memory, language, and human judgement operating under pressure.
TARA doesn't try to eliminate that. Bias is not a bug in human cognition — it is architecture. What TARA does is listen to the conversation as it happens, surface patterns that suggest bias may be shaping the narrative, and hand that signal to a human who can decide what to do with it. The goal is always the same: give the person who matters most — the one being reviewed — a fairer chance.
“The goal isn't bias-free reviews. There's no such thing. The goal is fairer reviews — where bias gets caught and named before it hardens into a number.
— TalentSpotify
What TARA listens for
Biases are grouped into three families by the type of distortion they introduce.
How memory, framing, and mental shortcuts distort the rating
How identity, power, and similarity shape who gets the benefit of the doubt
How scale usage drifts when managers avoid the extremes
What it is
The whole review period collapses into its final few weeks. Whatever happened most recently feels the most true — so a strong finish papers over a weak stretch, or one late stumble erases ten strong months.
Scenario
Sounds like
“She's been a bit shaky lately, honestly.”
The fairer move
What it is
The opposite anchor. A first impression — good or bad — sets a frame that later evidence struggles to shift. The opening read quietly becomes the lens for everything after it.
Scenario
Sounds like
“He took a while to find his feet — that's kind of who he is.”
The fairer move
What it is
One genuine strength spills over into unrelated areas. Because someone is excellent at one visible thing, they get the benefit of the doubt everywhere — without separate evidence.
Scenario
Sounds like
“She's so impressive in the room, she must be on top of everything.”
The fairer move
What it is
The mirror of the halo. One weakness darkens the entire evaluation, eclipsing the real strengths the person actually has.
Scenario
Sounds like
“After that miss, I just can't fully trust his output.”
The fairer move
What it is
A number or label from before — last cycle's rating, a first guess — becomes the gravitational centre. The review nudges slightly around the anchor instead of starting fresh from this period's evidence.
Scenario
Sounds like
“She's always been around a 3, so… maybe a 3-plus?”
The fairer move
What it is
The conclusion comes first; the evidence-gathering then quietly selects for examples that fit it and skips the ones that don't.
Scenario
Sounds like
“Every example I can think of points the same way.”
The fairer move
What it is
Outcomes get pinned on character instead of circumstance. A miss becomes 'they're not driven' rather than 'the goalposts moved three times' — and the situation that actually shaped the result disappears.
Scenario
Sounds like
“It's a motivation thing with him, I think.”
The fairer move
What it is
Everyone gets rated into the safe middle. Avoiding the top and the bottom dodges hard conversations — and flattens the real differences between people.
Scenario
Sounds like
“I keep everyone around the middle — it's cleaner.”
The fairer move
What it is
Assumptions tied to gender shape how the same behaviour is read — praised in one person, penalised in another. Assertive becomes 'aggressive'; warm becomes 'soft.'
Scenario
Sounds like
“She comes across too strong with the team.”
The fairer move
What it is
Assumptions tied to age, cutting both ways — younger people seen as not-ready-yet, older people as set-in-their-ways. Capability gets read off a birth year instead of the work.
Scenario
Sounds like
“He's only 24 — I can't put him in front of the client.”
The fairer move
What it is
We rate people who remind us of ourselves more generously — same school, same background, same communication style. Comfort gets mistaken for competence.
Scenario
Sounds like
“He just gets it — we think the same way.”
The fairer move
What it is
Power gaps distort the conversation. Rank substitutes for evidence, or a senior voice flattens a junior one — so the review stops being a two-way exchange.
Scenario
Sounds like
“Let's not overthink this — I've decided how it went.”
The fairer move
What it is
Ratings drift upward to avoid friction. It feels kind, but it quietly denies people the honest signal they need to grow — and erodes the meaning of the scale for everyone else.
Scenario
Sounds like
“I'll mark it a 4 — it's easier than getting into it.”
The fairer move
What it is
The mirror image — a bar set so high almost nobody clears it. Reads as rigour; lands as discouragement, and makes genuinely strong work look mediocre.
Scenario
Sounds like
“Nobody on my team gets a 5 — that's just my standard.”
The fairer move
And one thing TARA does not score: harmful language.
Aggressive, demeaning or dismissive wording is not treated as bias and never touches the score. It routes straight to a separate human safety review. A “we should just let him go” said mid-review is a people-risk event in its own right — whether or not any bias scored high. A number should never decide whether words crossed a line.
From signal to score, transparently
Every bias TARA detects produces a signal. Signals combine into one auditable number. Here is how.
Detection and scoring stay separate
The AI detects
- Which bias pattern is present
- The exact quote triggering the signal
- A plain-language reason
- How confident it is
never mixes
The system decides
- Aggregate the signals
- Apply fixed severity weights
- Normalise to 0–100
- Route to the right action
How the score works
Every signal TARA detects is evaluated against multiple factors — including how clearly the bias is present in the language, the inherent risk level of that bias type, and whether it appeared more than once in the conversation. These factors are combined into a single number between 0 and 100.
Weighted Bias Score — one number, 0 to 100.
The exact calculation is proprietary and not published. What is published: the evidence behind every flag, and the band it lands in.
The result is a single number from 0 to 100. That number maps to one of five action bands — each with a defined, automatic response.
But the score is only half the story. Context determines where the band threshold sits. The same score can land in different bands depending on whether a protected characteristic was involved — because the stakes are different, not because the evidence changed.
Standard context
The same score lands in the Coaching band under standard settings.
Protected-category context
With a protected characteristic detected, the same score routes to HR Review instead.
Two hardcoded floors ensure that certain signals can never be underweighted, regardless of how the rest of the conversation scored.
Protected-category floor
↑ Only escalates, never downgrades
When a bias touching a protected characteristic (gender, age) is detected, the score can only go up from the floor — not be averaged away by lower signals elsewhere.
Safety lane
↑ Bypasses scoring entirely
Harmful, demeaning, or dismissive language is not scored as bias. It routes straight to a separate human safety review — a number should never decide whether words crossed a line.
Putting it all together: here is how a real review conversation — with three bias signals and a floor override — resolves into an outcome.
Illustrative example · numbers simplified to show the logic
Three signals surface
Score computed
WBS · Awareness band
Would normally trigger: manager nudge
Protected-category floor fires
↑ Floor override: Gender / Tone detected at high confidence. Score can only escalate from the protected-category floor — Awareness band is overridden.
Outcome
Routed to HR Review — not ‘Awareness.’
The floor, not the aggregate score, determined the action.
Built to be questioned
The score is only trustworthy if the system behind it is. These five rails are not features — they are the conditions under which TARA is allowed to operate.
A signal, never a verdict
Every output carries it in plain words. TARA points to evidence; it doesn't pass judgement on a person.
A human decides, always
Nothing TARA produces touches an employee record on its own. A person reviews every flag before any action follows.
Protected categories are floored, on by default
Gender, age and other protected characteristics can't be scored away. The escalation floor ships switched on.
Every score is reproducible
Input → settings → score → action is fully logged and can be replayed on demand. The same evidence always yields the same result.
India-first by design
Built around the principles of India's Digital Personal Data Protection (DPDP) Act, with consent and auditability treated as features, not afterthoughts.
Fairer reviews start with hearing the bias out loud
See TARA in a live conversation — real signals, real language, one auditable score.
See TARA in a live conversationTARA surfaces bias signals to support human judgement. It reduces bias signals; it does not eliminate bias. This is a signal, not a verdict.